Construção de redes complexas de interações entre SNPs: uma abordagem computacional e estatística

Abstract

Single nucleotide polymorphisms (SNPs) represent the most common form of genetic variation in the human genome and have been extensively studied due to their impact on phenotypic traits and susceptibility to complex diseases. Although genome-wide association studies (GWAS) have yielded important discoveries by testing each SNP marginally, this approach tends to overlook dependencies between loci, a phenomenon known as epistasis, which may leave part of the phenotypic genetic variance unexplained. To address this limitation, the present work proposes a two-step approach for identifying SNP–SNP interactions associated with blood glucose levels (mg/dL). In the first step, simple linear regression models are fitted to pre-select SNPs with the strongest evidence of marginal association with the phenotype. In the second step, Interaction Forests are applied to the selected subset to rank SNP pairs according to their interaction effect, as quantified by the Effect Importance Measure (EIM). Prioritized pairs are subsequently validated using linear regression models with interaction terms. Missing SNPs were imputed using Markov chains and recombination fractions, and the data were obtained from the ISA-Capital study, comprising 698 genotyped participants and 324,471 SNPs after quality control. Simulation results demonstrated that Interaction Forests consistently recovered pairs with artificially introduced interactions under both quantitative and qualitative scenarios. In the application to real data, the regression-based pre-selection strategy (Case 3) yielded the lowest interaction p-values among the three approaches evaluated, with pairs reaching p < 10^-15, and enabled the construction of SNP interaction networks with connected components and identifiable cycles. These findings reinforce the potential of the proposed methodology for detecting epistasis in high-dimensional genomic data.